[PATCH 1/1] Per socket value for max datagram queue length

[PATCH 1/1] Per socket value for max datagram queue length

[Dan Ballard]

Provides a new option for setsockopt SO_MAX_DGRAM_QLEN that sets and
gets a socket specific max datagram queue length. Currently each socket
has one but it's only ever initialized from
/proc/sys/net/unix/max_dgram_qlen and then never adjustable later. Now
each socket can have it individually tweaked during it's life.

Signed-off-by: Dan Ballard <dan@mindstab.net>
---
  include/uapi/asm-generic/socket.h |    2 ++
  net/core/sock.c                   |    7 +++++++
  2 files changed, 9 insertions(+)

diff --git a/include/uapi/asm-generic/socket.h 
b/include/uapi/asm-generic/socket.h
index 38f14d0..f8c3e6b 100644
--- a/include/uapi/asm-generic/socket.h
+++ b/include/uapi/asm-generic/socket.h
@@ -80,4 +80,6 @@

  #define SO_MAX_PACING_RATE     47

+#define SO_MAX_DGRAM_QLEN      48
+
  #endif /* __ASM_GENERIC_SOCKET_H */
diff --git a/net/core/sock.c b/net/core/sock.c
index 5393b4b..1ff69d1 100644
--- a/net/core/sock.c
+++ b/net/core/sock.c
@@ -915,6 +915,10 @@ set_rcvbuf:
                                          sk->sk_max_pacing_rate);
                 break;

+       case SO_MAX_DGRAM_QLEN:
+               sk->sk_max_ack_backlog = val;
+               break;
+
         default:
                 ret = -ENOPROTOOPT;
                 break;
@@ -1182,6 +1186,9 @@ int sock_getsockopt(struct socket *sock, int 
level, int optname,
                 v.val = sk->sk_max_pacing_rate;
                 break;

+       case SO_MAX_DGRAM_QLEN:
+               v.val = sk->sk_max_ack_backlog;
+               break;
         default:
                 return -ENOPROTOOPT;
         }
-- 
1.7.10.4

Re: [PATCH 1/1] Per socket value for max datagram queue length

[Hannes Frederic Sowa]

On Wed, Jan 22, 2014 at 07:11:20AM -0800, Dan Ballard wrote:
> diff --git a/net/core/sock.c b/net/core/sock.c
> index 5393b4b..1ff69d1 100644
> --- a/net/core/sock.c
> +++ b/net/core/sock.c
> @@ -915,6 +915,10 @@ set_rcvbuf:
>                                          sk->sk_max_pacing_rate);
>                 break;
> 
> +       case SO_MAX_DGRAM_QLEN:
> +               sk->sk_max_ack_backlog = val;
> +               break;
> +

Shouldn't the backlog be capped for unprivileged users to some configurable
value? I even think that max_dgram_qlen should be the upper bound.

I guess it is not that serious as socket read accounting does account all
packets which sit in the backlog queue.

Greetings,

  Hannes

Re: [PATCH 1/1] Per socket value for max datagram queue length

[Eric Dumazet]

On Wed, 2014-01-22 at 07:11 -0800, Dan Ballard wrote:
> Provides a new option for setsockopt SO_MAX_DGRAM_QLEN that sets and
> gets a socket specific max datagram queue length. Currently each socket
> has one but it's only ever initialized from
> /proc/sys/net/unix/max_dgram_qlen and then never adjustable later. Now
> each socket can have it individually tweaked during it's life.
> 

Your patch suffers from many problems.

It breaks listen()

It doesn't compile on a lot of arches.

sk_max_ack_backlog is an unsigned short.

max_dgram_qlen is used, but only for Unix sockets at this time
Try : "git grep -n max_dgram_qlen" for details


> Signed-off-by: Dan Ballard <dan@mindstab.net>
> ---
>   include/uapi/asm-generic/socket.h |    2 ++
>   net/core/sock.c                   |    7 +++++++
>   2 files changed, 9 insertions(+)
> 
> diff --git a/include/uapi/asm-generic/socket.h 
> b/include/uapi/asm-generic/socket.h
> index 38f14d0..f8c3e6b 100644
> --- a/include/uapi/asm-generic/socket.h
> +++ b/include/uapi/asm-generic/socket.h
> @@ -80,4 +80,6 @@
> 
>   #define SO_MAX_PACING_RATE     47
> 
> +#define SO_MAX_DGRAM_QLEN      48
> +
>   #endif /* __ASM_GENERIC_SOCKET_H */
> diff --git a/net/core/sock.c b/net/core/sock.c
> index 5393b4b..1ff69d1 100644
> --- a/net/core/sock.c
> +++ b/net/core/sock.c
> @@ -915,6 +915,10 @@ set_rcvbuf:
>                                           sk->sk_max_pacing_rate);
>                  break;
> 
> +       case SO_MAX_DGRAM_QLEN:
> +               sk->sk_max_ack_backlog = val;
> +               break;
> +
>          default:
>                  ret = -ENOPROTOOPT;
>                  break;
> @@ -1182,6 +1186,9 @@ int sock_getsockopt(struct socket *sock, int 
> level, int optname,
>                  v.val = sk->sk_max_pacing_rate;
>                  break;
> 
> +       case SO_MAX_DGRAM_QLEN:
> +               v.val = sk->sk_max_ack_backlog;
> +               break;
>          default:
>                  return -ENOPROTOOPT;
>          }

Re: [PATCH 1/1] Per socket value for max datagram queue length

[Daniel Borkmann]

On 01/22/2014 04:11 PM, Dan Ballard wrote:
> Provides a new option for setsockopt SO_MAX_DGRAM_QLEN that sets and
> gets a socket specific max datagram queue length. Currently each socket
> has one but it's only ever initialized from
> /proc/sys/net/unix/max_dgram_qlen and then never adjustable later. Now
> each socket can have it individually tweaked during it's life.
>
> Signed-off-by: Dan Ballard <dan@mindstab.net>
> ---
>   include/uapi/asm-generic/socket.h |    2 ++
>   net/core/sock.c                   |    7 +++++++
>   2 files changed, 9 insertions(+)
>
> diff --git a/include/uapi/asm-generic/socket.h b/include/uapi/asm-generic/socket.h
> index 38f14d0..f8c3e6b 100644
> --- a/include/uapi/asm-generic/socket.h
> +++ b/include/uapi/asm-generic/socket.h
> @@ -80,4 +80,6 @@
>
>   #define SO_MAX_PACING_RATE     47
>
> +#define SO_MAX_DGRAM_QLEN      48
> +

This needs to be added in more than just asm-generic, e.g.
have a look at SO_MAX_PACING_RATE or SO_BPF_EXTENSIONS.

Also you might need to rebase to current net-next head and
maybe describe use cases more in-depth; next to what Hannes
just commented.

>   #endif /* __ASM_GENERIC_SOCKET_H */
> diff --git a/net/core/sock.c b/net/core/sock.c
> index 5393b4b..1ff69d1 100644
> --- a/net/core/sock.c
> +++ b/net/core/sock.c
> @@ -915,6 +915,10 @@ set_rcvbuf:
>                                           sk->sk_max_pacing_rate);
>                  break;
>
> +       case SO_MAX_DGRAM_QLEN:
> +               sk->sk_max_ack_backlog = val;
> +               break;
> +
>          default:
>                  ret = -ENOPROTOOPT;
>                  break;
> @@ -1182,6 +1186,9 @@ int sock_getsockopt(struct socket *sock, int level, int optname,
>                  v.val = sk->sk_max_pacing_rate;
>                  break;
>
> +       case SO_MAX_DGRAM_QLEN:
> +               v.val = sk->sk_max_ack_backlog;
> +               break;
>          default:
>                  return -ENOPROTOOPT;
>          }

Re: [PATCH 1/1] Per socket value for max datagram queue length

[Hannes Frederic Sowa]

On Wed, Jan 22, 2014 at 04:20:36PM +0100, Hannes Frederic Sowa wrote:
> On Wed, Jan 22, 2014 at 07:11:20AM -0800, Dan Ballard wrote:
> > diff --git a/net/core/sock.c b/net/core/sock.c
> > index 5393b4b..1ff69d1 100644
> > --- a/net/core/sock.c
> > +++ b/net/core/sock.c
> > @@ -915,6 +915,10 @@ set_rcvbuf:
> >                                          sk->sk_max_pacing_rate);
> >                 break;
> > 
> > +       case SO_MAX_DGRAM_QLEN:
> > +               sk->sk_max_ack_backlog = val;
> > +               break;
> > +
> 
> Shouldn't the backlog be capped for unprivileged users to some configurable
> value? I even think that max_dgram_qlen should be the upper bound.
> 
> I guess it is not that serious as socket read accounting does account all
> packets which sit in the backlog queue.

Just a follow-up:

sk_max_ack_backlog is also responsible for limiting the af_unix
dgram queues.  Currently there is no socket accounting for the read
side of those unix dgram sockets. I tried to fix this once here,
http://patchwork.ozlabs.org/patch/231032/, but until that is done we
depend on max_dgram_qlen to limit those queues at all.

I hope I can get back to this patch anytime soon, as it solves the problem
that a bidirectional protocol ping-ponging with a dgram server socket
and not fetching its messages from the backlog queue can bring a server
to halt because it doesn't have any send space on the socket anymore.

Greetings,

  Hannes